Deep Hybrid Architecture for Suspicious Action Detection in Video Surveillance

For Ensuring safety for the public and their property, detection of suspicious actions from video surveillance is an important application in the rapid development of computer vision technology. In this paper, we propose a Deep Hybrid architecture for detecting suspicious activities in surveillance video. In this case, there are four stages in identifying abnormal actions: preprocessing, segmentation, feature extraction and detection. Initially, the video undergoes the preprocessing stage, here the video is converted into frames the Gaussian filtering method is applied to eliminate noise. Subsequently, frames are segmented using a deep joint segmentation model. This process takes place in the segmentation stage. Then features like improved LGXP and SLBT features are extracted in the feature extraction stage. Modified Gabor filter is employed for the extraction process of the improved LGXP feature. The hybrid classification model is used for identifying suspicious actions. The combined SLBT and improved LGXP features are given as input to the hybrid classification model. This model is a combination of CNN and Bi-GRU frameworks. The detection of abnormal actions is by averaging the outcome of CNN and Bi-GRU. The model detects the suspicious activities like traffic collisions, attacks, mistreatment and gun firing.